CN114610871A - Information system modeling analysis method based on artificial intelligence algorithm - Google Patents

Information system modeling analysis method based on artificial intelligence algorithm Download PDF

Info

Publication number
CN114610871A
CN114610871A CN202210511565.0A CN202210511565A CN114610871A CN 114610871 A CN114610871 A CN 114610871A CN 202210511565 A CN202210511565 A CN 202210511565A CN 114610871 A CN114610871 A CN 114610871A
Authority
CN
China
Prior art keywords
encoder
hidden state
decoder
moment
vector
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210511565.0A
Other languages
Chinese (zh)
Other versions
CN114610871B (en
Inventor
张晟
杨晓冬
王吉平
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Daoda Tianji Technology Co ltd
Original Assignee
Beijing Daoda Tianji Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Daoda Tianji Technology Co ltd filed Critical Beijing Daoda Tianji Technology Co ltd
Priority to CN202210511565.0A priority Critical patent/CN114610871B/en
Publication of CN114610871A publication Critical patent/CN114610871A/en
Application granted granted Critical
Publication of CN114610871B publication Critical patent/CN114610871B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/34—Browsing; Visualisation therefor
    • G06F16/345—Summarisation for human users
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A10/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE at coastal zones; at river basins
    • Y02A10/40—Controlling or monitoring, e.g. of flood or hurricane; Forecasting, e.g. risk assessment or mapping

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Databases & Information Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Machine Translation (AREA)

Abstract

The invention relates to an information system modeling analysis method based on an artificial intelligence algorithm, which comprises the following steps: analyzing the model document by using a neural language model to obtain a plurality of word vectors; performing chapter decomposition on the model document to obtain a model full text and n chapters corresponding to the model document, wherein the model full text and the n chapters both comprise a plurality of word vectors; processing the model full text through an LSTM recurrent neural network to generate a full text thought vector c; processing the n chapters through an LSTM recurrent neural network based on the full-text thought vector c to generate chapter thought vectors cp; and processing the chapter thought vector cp through an LSTM recurrent neural network to generate a model abstract and realize intention analysis. The LSTM recurrent neural network used in the scheme has correct intention analysis, automatically processes and generates the model abstract, not only can improve the accuracy of information analysis results, but also can reduce the time and energy cost of information analysts.

Description

Information system modeling analysis method based on artificial intelligence algorithm
Technical Field
The invention relates to the technical field of information analysis, in particular to an information system modeling analysis method based on an artificial intelligence algorithm.
Background
With the development of information technology, a large number of new technologies are applied to the field of intelligence analysis, and the intelligence gathering capacity is greatly improved. Intelligence is a prerequisite and basis for decision making, but not much of an advantage, as things can go the opposite if a certain margin is exceeded, especially today with information inundation. Various information of the internet is covered on the ground, the truth is difficult to distinguish, the information cannot be directly used without analysis, and valuable information can be obtained only by analysis.
The intelligence analysis is a process of information selection and synthesis, an analyst acquires data through a conventional search engine by adopting a traditional analysis mode, manages the data by a mode of manually arranging documents and the like, and consumes a large amount of time and energy, and an intelligence analysis model accumulated in the process cannot be directly used in the process of information collection and analysis, so that the intelligence analysis effect is not ideal.
Various models are widely used and applied in the field of information analysis, and the information analysis models are more and more computerized, intelligent, full-resource and modeled. How to really apply the information analysis model to an actual information system to realize automation and intellectualization is difficult, and because the business modeling of the information has no certain rule and can not be structured, the information analysis model generally adopts natural language. However, the intelligence analysis model currently used faces the following problems when performing intelligence gathering and analysis:
How to make a computer read an information analysis model and accurately comprehend the intention of a modeler. To realize intelligence collection and analysis, the model must be analyzed intentionally, and the accuracy of the analysis determines the accuracy of the final intelligence analysis result.
And secondly, analyzing how the searched information is, so that the information can be close to the requirement of the original model to the maximum extent, and further generating an analysis report. The information technology has great promotion effect on information collection and analysis, for example, keywords and the like can be used for quick search and analysis, the existing information search system can absorb all information data in an undesirable way, and huge consumption of time and energy is caused to information analysts.
Therefore, it is necessary to improve the accuracy of the information analysis result and to reduce the time and effort cost of the information analyst.
Disclosure of Invention
The invention aims to improve the accuracy of information analysis results and reduce the time and energy costs of information analysts, and provides an information system modeling analysis method based on an artificial intelligence algorithm.
In order to achieve the above object, the embodiments of the present invention provide the following technical solutions:
An intelligence system modeling analysis method based on an artificial intelligence algorithm comprises the following steps:
step S1: collecting and analyzing intelligence by using an intelligence analysis model so as to generate a model document;
step S2: analyzing the model document by using a neural language model trained based on a domain corpus, thereby obtaining a plurality of Word vectors, wherein the neural language model is a Word2Vec model; performing chapter decomposition on the model document to obtain a model full text and n chapters corresponding to the model document, wherein the model full text and the n chapters respectively comprise a plurality of word vectors;
step S3: processing the model full text through an LSTM recurrent neural network to generate a full text thought vector c; processing the n chapters through an LSTM recurrent neural network based on the full-text thought vector c to generate chapter thought vectors cp;
step S4: and processing the chapter thought vector cp through an LSTM recurrent neural network to generate a model abstract and realize intention analysis.
In the scheme, the existing information analysis model is used for acquiring information, but the acquired information is disordered and has good and bad information, so the scheme processes the acquired information of the information analysis model, and the processing process comprises the steps of firstly decomposing the information (namely a model document) into a plurality of word vectors, then automatically processing the word vectors through an LSTM recurrent neural network to generate a model abstract, and obtaining the key information from the model abstract. The LSTM recurrent neural network has correct intention analysis, automatically processes and generates the model abstract, not only can improve the accuracy of information analysis results, but also can reduce the time and energy cost of information analysts.
The step S1 specifically includes the following steps: the information analysis model is utilized to establish task nodes according to the analysis requirements of users, working contents are set for the nodes established by each information analysis model, and the working contents can be search contents, search ranges, analysis methods and the like, so that an analysis link is formed, and the analysis link is a model document.
The step S3 is preceded by the steps of: training the LSTM recurrent neural network:
the LSTM recurrent neural network comprises an encoder and a decoder;
sequentially inputting word vector training set X = { X) to encoder of LSTM recurrent neural networktIs formed by T ∈ N, N is an integer which is more than or equal to 1, and xtRepresenting the word vector input to the encoder at time t; when t =1, a first initial hidden state h is input to the encoder0And x1The encoder outputs the hidden state h at this moment1(ii) a At t>When 1, the hidden state h of the last moment is input into the encodert-1And the word vector x at this momenttThe encoder outputs the hidden state h at this momentt(ii) a Up to all word vectors xtAfter all the input encoders are finished, the encoder finally outputs the hidden state h of the Nth momentN；
Sequentially inputting a predictive value training set Y = { Y ] to a decoder of the LSTM recurrent neural networkt`T' is an integer greater than or equal to 1, M is an integer greater than or equal to 1, y t`Represents the prediction value input to the decoder at time t'; when t '= 1, a second initial hidden state h' is input to the decoder0And initial predicted value y1The second initial hidden state h ″0Outputting a hidden state h of the Nth moment for the encoderNInitial predicted value y1Is a self-defined value<bos>The decoder outputs the hidden state h' at this moment1And the predicted value y of the next time2(ii) a At t ″)>At 1, the hidden state h' of the last time is input to the decodert`-1And the predicted value y at that timet`The decoder outputs the concealment at this momentState h' oft`And the predicted value y at the next timet`+1(ii) a Until all the predicted values yt`All input to the decoder, or until the decoder outputs the custom predictor yt`+1=<eos>；
Thereby obtaining a trained encoder and a trained decoder.
In the above scheme, a large number of word vectors are collected as a training set to train the encoder and decoder of the LSTM recurrent neural network, so that the LSTM recurrent neural network can have correct intent resolution.
The step of processing the model full text through the LSTM recurrent neural network to generate a full text thought vector c comprises the following steps:
the LSTM recurrent neural network comprises a trained encoder and a trained decoder;
the model contains T word vectors X in the whole text, and the word vector X = { X = tInputting the codes into the coder in sequence, T belongs to T, T is an integer which is more than or equal to 1, xtRepresents the word vector input to the encoder at time t; when t =1, a first initial hidden state h is input to the trained encoder0Word vector x1The encoder outputs the hidden state h at this moment1(ii) a When t is>When 1, the hidden state h of the last moment is input into the encodert-1And the word vector x at this momenttThe encoder outputs the hidden state h at this momentt(ii) a Up to all word vectors xtAll input into the encoder to obtain the hidden state h at the Tth momentT；
Retiming, predicting word Y = { Y = }t`Is sequentially input into a decoder, T' is an integer greater than or equal to 1, yt`Represents the prediction value input to the decoder at time t'; when t '= 1, inputting a second initial hidden state h' into the trained decoder0And the initial predicted value y1The second initial hidden state h ″, the first initial hidden state0Hidden state h for encoder outputTInitial predicted value y1Is a custom value<bos>The decoder outputs the hidden state h' at this moment1And the predicted value y at the next time2(ii) a At t ″)>At 1, the hidden state h' of the last time is input to the decodert`-1And the predicted value y at that timet`The decoder outputs the hidden state h' at this momentt`And the predicted value y at the next timet`+1(ii) a Until T' is reached, or until the decoder outputs the custom prediction value y t`+1=<eos>；
And generating a full-text thought vector c according to the hidden states output by the encoder and the decoder.
In the scheme, the trained LSTM recurrent neural network is used for processing the full text of the model to generate a full text thought vector c.
The step of generating the full-text thought vector c according to the hidden states output by the encoder and the decoder comprises the following steps: at the t' time of decoder, full text thought vector c can be generatedt`：
Figure DEST_PATH_IMAGE001
Wherein, the first and the second end of the pipe are connected with each other,
Figure DEST_PATH_IMAGE002
full text thought vector c representing time tt`I is the ith time of the encoder, hiA hidden state output for the ith moment of the encoder;
weight of
Figure 962002DEST_PATH_IMAGE002
The calculation of (2):
hidden state h 'output by decoder at t' momentt`And the hidden state h output by the encoder at each momentiCalculating a score
Figure DEST_PATH_IMAGE003
(ii) a After the score is processed by softmax, the score is processed by softmax
Figure DEST_PATH_IMAGE004
Is converted into
Figure 941459DEST_PATH_IMAGE002
；
Obtaining a full-text thought vector c = { c = }t`}，t`∈T`。
The hidden state h 'output by the decoder at the t' momentt`And the hidden state h output by the encoder at each momentiCalculating a score
Figure 504552DEST_PATH_IMAGE003
Comprises the following steps:
Figure DEST_PATH_IMAGE005
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE006
、
Figure DEST_PATH_IMAGE007
、
Figure DEST_PATH_IMAGE008
representing learnable parameters, h ″t`And hiThe combined signals are input into a multi-layer perceptron to obtain a score, and tanh is an activation function.
The step of processing the n chapters based on the full-text thought vector c through the LSTM recurrent neural network to generate the chapter thought vector cp comprises the following steps:
The LSTM recurrent neural network comprises a trained encoder and a trained decoder;
analyzing n chapters based on the full-text thought vector c, wherein the n chapters contain K word vectors X, and the word vectors X = { X =kK belongs to K, K is an integer which is more than or equal to 1, and xkRepresents the word vector input to the encoder at time k; when k =1, inputting a third initial hidden state g to the trained encoder0Sum word vector x1The encoder outputs the hidden shape g at this moment1(ii) a When k is>When 1, the hidden state g of the last time is input to the encoderk-1And the word vector x at this momentkThe encoder outputs the hidden state g at this momentk(ii) a Up to all word vectors xkAll input into the encoder, the hidden state g at the Kth moment is obtainedK；
Retiming, predicting word Y = { Y = }k`Inputting the data into a decoder in sequence, K' is an integer greater than or equal to 1, and yk`Represents the prediction value input to the decoder at time k'; when k =1, inputting a fourth initial hidden state g' into the trained decoder0And the initial predicted value y1The fourth initial latent state g ″0Hidden state g for encoder outputKInitial predicted value y1Is a custom value<bos>The decoder outputs the hidden state g' at this moment1And the predicted value y at the next time2(ii) a At k>At 1, the hidden state g' at the last moment is input to the decoder k`-1And the predicted value y at that timek`The decoder outputs the hidden state g' at this momentk`And the predicted value y at the next timek`-1(ii) a Until K' is reached, or until the decoder outputs the custom prediction value yk`+1=<eos>；
The chapter idea vector cp is generated from the hidden states output by the encoder and decoder.
In the scheme, the word vectors in n chapters are analyzed by using the full-text thought vector c, and then the word vectors in n chapters are processed by using the trained LSTM recurrent neural network to generate the chapter thought vector cp.
The step of generating the chapter idea vector cp according to the hidden states output by the encoder and the decoder includes: at time k' of the decoder, a chapter thought vector cp may be generatedk`：
Figure DEST_PATH_IMAGE009
Wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE010
shows the chapter idea vector cp at time k `k`J is the jth moment of the encoder, gjA hidden state output for the j moment of the encoder;
weight of
Figure 775128DEST_PATH_IMAGE010
The calculation of (2):
hidden state g 'output by decoder at k' momentk`And the hidden state g output by the encoder at each momentjCalculating a score
Figure DEST_PATH_IMAGE011
(ii) a After the score is processed by softmax, the score is processed by softmax
Figure DEST_PATH_IMAGE012
Is converted into
Figure 651817DEST_PATH_IMAGE010
；
Get chapter idea vector cp = { cp = { (cp)k`}，k`∈K`。
The hidden state g 'output at the k' th moment of the decoderk`And the hidden state g output by the encoder at each moment jCalculating a score
Figure 739859DEST_PATH_IMAGE011
Comprises the following steps:
Figure DEST_PATH_IMAGE013
wherein, the first and the second end of the pipe are connected with each other,
Figure 52897DEST_PATH_IMAGE006
、
Figure 517376DEST_PATH_IMAGE007
、
Figure 451834DEST_PATH_IMAGE008
representing a learnable parameter, g ″)k`And gjThe combined signals are input into a multi-layer perceptron to obtain a score, and tanh is an activation function.
The method comprises the following steps of processing chapter thought vectors cp through an LSTM recurrent neural network to generate a model abstract and realize intention analysis, and comprises the following steps of:
at the k' th moment of the decoder, the chapter thought vector is divided into twocpk`And hidden state g' output by decoderk`After splicing together, inputting a full connection layer, and obtaining an abstract through softmax:
Figure DEST_PATH_IMAGE014
where Pvocab represents the chapter idea vector cpk`Selecting the chapter thought vector cp with the maximum probability value at the moment k' according to the corresponding probability valuek`Chapter p as the time;
thereby generating a model section P = { P = { P }k`}，k`∈K`。
In the scheme, a final model abstract is obtained according to the generated chapter idea vector cp.
Compared with the prior art, the invention has the beneficial effects that:
the LSTM recurrent neural network used in the scheme has correct intention analysis, automatically processes and generates the model abstract, not only can improve the accuracy of information analysis results, but also can reduce the time and energy cost of information analysts.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the embodiments will be briefly described below, it should be understood that the following drawings only illustrate some embodiments of the present invention and therefore should not be considered as limiting the scope, and for those skilled in the art, other related drawings can be obtained according to the drawings without inventive efforts.
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a process for training the LSTM recurrent neural network according to an embodiment of the present invention;
FIG. 3 is a diagram illustrating the processing of the model full text using the LSTM recurrent neural network according to an embodiment of the present invention;
FIG. 4 is a process of using the LSTM recurrent neural network for n chapters according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. The components of embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the present invention, presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present invention without making any creative effort, shall fall within the protection scope of the present invention.
It should be noted that: like reference numbers and letters refer to like items in the following figures, and thus, once an item is defined in one figure, it need not be further defined and explained in subsequent figures. Also, in the description of the present invention, the terms "first", "second", and the like are used for distinguishing between descriptions and not necessarily for describing a relative importance or implying any actual relationship or order between such entities or operations.
Example 1:
the invention is realized by the following technical scheme, and the intelligence system modeling analysis method based on the artificial intelligence algorithm comprises the following steps:
step S1: intelligence gathering and analysis is performed using an intelligence analysis model, thereby generating a model document.
The method comprises the steps of establishing task nodes according to user analysis requirements by utilizing the existing information analysis model, setting working contents for the nodes established by each information analysis model, wherein the working contents can be search contents, search ranges, analysis methods and the like, so that an analysis link is formed, and the analysis link is a model document.
Step S2: analyzing the model document by using a neural language model trained based on a domain corpus so as to obtain a plurality of Word vectors, wherein the neural language model is a Word2Vec model; and performing chapter decomposition on the model document to obtain a model full text and n chapters corresponding to the model document, wherein the model full text and the n chapters respectively comprise a plurality of word vectors.
The model document is analyzed in the step, and the purpose is to decompose the model document into a model full text and n chapters, wherein the model full text consists of T word vectors, and the n chapters consist of K 'word vectors, so that the model document consists of T + K' word vectors. However, in step S3, n chapters composed of K 'word vectors are modified according to the generated full-text thought vector c, for example, K' word vectors contained in n chapters are modified into K word vectors, and the thought intent of the model full-text is added.
Step S3: processing the model full text through an LSTM recurrent neural network to generate a full text thought vector c; and processing the n chapters on the basis of the full-text thought vector c through an LSTM recurrent neural network to generate a chapter thought vector cp.
Before the model full text is processed using the LSTM recurrent neural network, which includes an encoder and a decoder, the LSTM recurrent neural network needs to be trained. During training, a large number of word vectors are collected, and the word vectors can be words commonly used in the industry, words commonly used in life, professional terms in the field and the like, and are not limited here.
Referring to fig. 2, a large number of word vectors form a training set X = { X = { [ X ]tN is an integer greater than or equal to 1, the N is the number of word vectors, the N word vectors in the training set X are input into an encoder of the LSTM recurrent neural network in sequence, the word vector input into the first encoder is X1I.e., t = 1; the last word vector input to the encoder is xNI.e. t = N, that is, xtRepresenting the word vector input to the encoder at time t.
When t =1, inputting a first initial implicit state h to the encoder0And x1The encoder outputs the hidden state h at this moment 1(ii) a At t>1 hour, the hidden state h of the previous moment is input into the encodert-1And the word vector x at this timetThe encoder outputs the hidden state h at this momentt(ii) a Up to all word vectors xtAfter all the input encoders are finished, the encoder finally outputs the hidden state h of the Nth momentN. It should be noted that the first initial hidden state h0The setting is customized for the user and can be set according to the actual situation.
Then, a predictive value training set Y = { Y ] is sequentially input to a decoder of the LSTM recurrent neural networkt`T' is an integer greater than or equal to 1, M is an integer greater than or equal to 1, yt`Indicating the predictor input to the decoder at time t', predictor yt`And the output of the decoder is obtained in turn. When t '= 1, inputting a second initial hidden state h' into a decoder0And the initial predicted value y1The second initial hidden state h ″, the first initial hidden state0Outputting a hidden state h of the Nth moment for the encoderNInitial predicted value y1Custom value for a user<bos>The decoder outputs the hidden state h' at this moment1And the predicted value y of the next time2(ii) a At t ″)>At 1, the hidden state h' of the last time is input to the decodert`-1And the predicted value y at that timet`The decoder outputs the hidden state h' at this momentt`And the predicted value y at the next timet`+1(ii) a Until all the predicted values yt`All input into the decoder and output the hidden state h' at the Mth time MAnd the predicted value y at the next timeM+1Or until the decoder outputs a custom predictor yt`+1=<eos>. It should be noted that the second initial hidden state h ″, is0The setting is customized for the user and can be set according to the actual situation.
Thereby obtaining a trained encoder and a trained decoder.
Referring to fig. 3, assuming that the model full text consists of T word vectors X, the word vector X = { X = { (X) }tInputting the codes into the coder in sequence, T belongs to T, T is an integer which is more than or equal to 1, xtRepresenting the word vector input to the encoder at time t; when t =1, a first initial hidden state h is input to the trained encoder0Word vector x1The encoder outputs the hidden state h at this moment1(ii) a When t is>When 1, the hidden state h of the last moment is input into the encodert-1And the word vector x at this momenttThe encoder outputs the hidden state at this momentState ht(ii) a Up to all word vectors xtAll input into the encoder to obtain the hidden state h at the Tth momentT。
Retiming, predicting word Y = { Y = }t`Is sequentially input into a decoder, T' is an integer greater than or equal to 1, yt`Represents the prediction value input to the decoder at time t'; when t '= 1, inputting a second initial hidden state h' into the trained decoder0And the initial predicted value y1The second initial hidden state h ″, the first initial hidden state 0Hidden state h for encoder outputTInitial predicted value y1Is a self-defined value<bos>The decoder outputs the hidden state h' at this moment1And the predicted value y at the next time2(ii) a At t>At 1, the hidden state h' of the last time is input to the decodert`-1And the predicted value y at that timet`The decoder outputs the hidden state h' at this momentt`And the predicted value y at the next timet`+1(ii) a Until T 'time is reached and the hidden state h' of the time is outputT`And the predicted value y at the next timeT`+1Or until the decoder outputs the custom prediction value yt`+1=<eos>。
Generating a full-text thought vector c according to the hidden state output by the encoder and the decoder, and generating the full-text thought vector c at the t' moment of the decodert`：
Figure 976357DEST_PATH_IMAGE001
Wherein the content of the first and second substances,
Figure 996265DEST_PATH_IMAGE002
full text thought vector c representing time tt`I is the ith time of the encoder, hiA hidden state output for the ith moment of the encoder;
weight of
Figure 998856DEST_PATH_IMAGE002
The calculation of (2):
hidden state h 'output by decoder at t' momentt`And the hidden state h output by the encoder at each momentiCalculating a score
Figure 804132DEST_PATH_IMAGE003
：
Figure 499556DEST_PATH_IMAGE005
Wherein the content of the first and second substances,
Figure 6761DEST_PATH_IMAGE006
、
Figure 547463DEST_PATH_IMAGE007
、
Figure 722093DEST_PATH_IMAGE008
representing learnable parameters, h ″t`And hiThe combined signals are input into a multi-layer perceptron to obtain a score, and tanh is an activation function.
After the score is processed by softmax, the score is processed by softmax
Figure 57259DEST_PATH_IMAGE004
Is converted into
Figure 51760DEST_PATH_IMAGE002
(ii) a Obtaining a full-text thought vector c = { c = }t`}，t`∈T`。
Referring to fig. 4, n chapters are analyzed using the full text thought vector c, assuming that n chapters are composed of K word vectors X, for which X = { X = { kThe processing is the same as that of the full-text model, K belongs to K, K is an integer which is more than or equal to 1, and xkRepresenting the word vector input to the encoder at time k.
When k =1, inputting a third initial hidden state g to the trained encoder0Sum word vector x1The encoder outputs the hidden shape g at this moment1(ii) a When k is>When 1, the hidden state g of the last time is input to the encoderk-1And the word vector x at this momentkThe encoder outputs the hidden state g at this momentk(ii) a Up to all wordsVector xkAll input into the encoder, the hidden state g at the Kth moment is obtainedK. In addition, the third initial hidden state g0The setting is customized for the user and can be set according to the actual situation.
Retiming, predicting word Y = { Y = }k`Inputting the data into a decoder in sequence, K' is an integer greater than or equal to 1, and yk`Represents the prediction value input to the decoder at time k'; when k =1, inputting a fourth initial hidden state g' into the trained decoder0And the initial predicted value y1The fourth initial latent state g ″0Hidden state g for encoder outputKInitial predicted value y1Is a custom value<bos>The decoder outputs the hidden state g' at this moment1And the predicted value y at the next time2(ii) a At k>At 1, the hidden state g' at the last moment is input to the decoder k`-1And the predicted value y at that timek`The decoder outputs the hidden state g' at this momentk`And the predicted value y at the next timek`-1(ii) a Until K 'is reached and the hidden state g' at this moment is outputK`And the predicted value y at the next timeK`-1Or until the decoder outputs the custom prediction value yk`+1=<eos>. It should be noted that the fourth initial hidden state g ″, is0The setting is customized for the user and can be set according to the actual situation.
The chapter thought vector cp is generated according to the hidden state output by the encoder and the decoder, and the chapter thought vector cp can be generated at the k' time of the decoderk`：
Figure 732290DEST_PATH_IMAGE009
Wherein the content of the first and second substances,
Figure 495846DEST_PATH_IMAGE010
shows the chapter idea vector cp at time k `k`J is the jth moment of the encoder, gjA hidden state output for the j moment of the encoder;
weight of
Figure 533072DEST_PATH_IMAGE010
The calculation of (2):
hidden state g 'output by decoder at k' momentk`And the hidden state g output by the encoder at each momentjCalculating a score
Figure 749290DEST_PATH_IMAGE011
：
Figure 897375DEST_PATH_IMAGE013
Wherein the content of the first and second substances,
Figure 515438DEST_PATH_IMAGE006
、
Figure 723565DEST_PATH_IMAGE007
、
Figure 177811DEST_PATH_IMAGE008
representing learnable parameters, g ″k`And gjThe combined signals are input into a multi-layer perceptron to obtain a score, and tanh is an activation function.
After the score is processed by softmax, the score is processed by softmax
Figure 864008DEST_PATH_IMAGE012
Is converted into
Figure 602157DEST_PATH_IMAGE010
(ii) a Get chapter idea vector cp = { cp = { (cp)k`}，k`∈K`。
Step S4: and processing the chapter thought vector cp through an LSTM recurrent neural network to generate a model abstract and realize intention analysis.
At the k' th time of the decoder, chapter idea vector cp is addedk`And hidden state g' output by decoderk`After splicing together, inputting a full connection layer, and obtaining an abstract through softmax:
Figure 981185DEST_PATH_IMAGE014
wherein Pvocab represents a chapter thought vector cpk`Selecting the chapter thought vector cp with the maximum probability value at the moment k' according to the corresponding probability valuek`Chapter p as the time;
thereby generating a model section P = { P = { P }k`}，k`∈K`。
The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (9)

1. An information system modeling analysis method based on artificial intelligence algorithm is characterized in that: the method comprises the following steps:
step S1: collecting and analyzing intelligence by using an intelligence analysis model so as to generate a model document;
step S2: analyzing the model document by using a neural language model trained based on a domain corpus so as to obtain a plurality of Word vectors, wherein the neural language model is a Word2Vec model; performing chapter decomposition on the model document to obtain a model full text and n chapters corresponding to the model document, wherein the model full text and the n chapters respectively comprise a plurality of word vectors;
Step S3: processing the model full text through an LSTM recurrent neural network to generate a full text thought vector c; processing the n chapters through an LSTM recurrent neural network based on the full-text thought vector c to generate chapter thought vectors cp;
step S4: and processing the chapter thought vectors cp through an LSTM recurrent neural network to generate a model abstract and realize intention analysis.
2. The intelligence system modeling analysis method based on artificial intelligence algorithm of claim 1, characterized in that: the step S1 specifically includes the following steps: the information analysis model is utilized to establish task nodes according to the analysis requirements of users, working contents are set for the nodes established by each information analysis model, and the working contents can be search contents, search ranges, analysis methods and the like, so that an analysis link is formed, and the analysis link is a model document.
3. The intelligence system modeling analysis method based on artificial intelligence algorithm of claim 1, characterized in that: the step S3 is preceded by the steps of: training the LSTM recurrent neural network:
the LSTM recurrent neural network comprises an encoder and a decoder;
sequentially inputting word vector training set X = { X) to encoder of LSTM recurrent neural network tIs formed by T ∈ N, N is an integer which is more than or equal to 1, and xtRepresenting the word vector input to the encoder at time t; when t =1, a first initial hidden state h is input to the encoder0And x1The encoder outputs the hidden state h at this moment1(ii) a At t>When 1, the hidden state h of the last moment is input into the encodert-1And the word vector x at this momenttThe encoder outputs the hidden state h at this momentt(ii) a Up to all word vectors xtAfter all the input encoders are finished, the encoder finally outputs the hidden state h of the Nth momentN；
Sequentially inputting a predictive value training set Y = { Y ] to a decoder of the LSTM recurrent neural networkt`T' is an integer greater than or equal to 1, M is an integer greater than or equal to 1, yt`Represents the prediction value input to the decoder at time t'; when t '= 1, inputting a second initial hidden state h' into a decoder0And the initial predicted value y1The second initial hidden state h ″, the first initial hidden state0Outputting a hidden state h of the Nth moment for the encoderNInitial predicted value y1Is a custom value<bos>The decoder outputs the hidden state h' at this moment1And the predicted value y of the next time2(ii) a At t ″)>At 1, the hidden state h' of the last time is input to the decodert`-1And the predicted value y at that timet`The decoder outputs the hidden state h' at this momentt`And the predicted value y at the next timet`+1(ii) a Until all the predicted values y t`All input to the decoder, or until the decoder outputs a custom predictor yt`+1=<eos>；
Thereby obtaining a trained encoder and a trained decoder.
4. The intelligence system modeling analysis method based on artificial intelligence algorithm of claim 1, wherein: the step of processing the model full text through the LSTM recurrent neural network to generate a full text thought vector c comprises the following steps:
the LSTM recurrent neural network comprises a trained encoder and a trained decoder;
the model contains T word vectors X in the whole text, and the word vector X = { X =tInputting into the coder in turn, T belongs to T, T is an integer greater than or equal to 1, xtRepresenting the word vector input to the encoder at time t; when t =1, a first initial hidden state h is input to the trained encoder0Word vector x1The encoder outputs the hidden state h at this moment1(ii) a When t is>When 1, the hidden state h of the last moment is input into the encodert-1And the word vector x at this momenttThe encoder outputs the hidden state h at this momentt(ii) a Up to all word vectors xtAll input into the encoder to obtain the hidden state h at the Tth momentT；
Retiming, predicting word Y = { Y = }t`Is sequentially input into a decoder, T' is an integer greater than or equal to 1, y t`Represents the prediction value input to the decoder at time t'; when t '= 1, a second initial hidden state h' is input to the trained decoder0And initial predicted value y1The second initial hidden state h ″0Hidden state h for encoder outputTInitial predicted value y1Is a custom value<bos>The decoder outputs the hidden state h' at this moment1And the predicted value y at the next time2(ii) a At t ″)>At 1, the hidden state h' of the last time is input to the decodert`-1And the predicted value y at that timet`The decoder outputs the implicit state at this momentState h' oft`And the predicted value y at the next timet`+1(ii) a Until T' is reached, or until the decoder outputs the custom prediction value yt`+1=<eos>；
And generating a full-text thought vector c according to the hidden states output by the encoder and the decoder.
5. The intelligence system modeling analysis method based on artificial intelligence algorithm of claim 4, characterized in that: the step of generating the full-text thought vector c according to the hidden states output by the encoder and the decoder comprises the following steps: at time t' of the decoder, a full-text thought vector c can be generatedt`：
Figure 123974DEST_PATH_IMAGE001
Wherein the content of the first and second substances,
Figure 509956DEST_PATH_IMAGE002
full text thought vector c representing time tt`I is the ith time of the encoder, hiA hidden state output for the ith moment of the encoder;
weight of
Figure 758535DEST_PATH_IMAGE002
The calculation of (2):
Hidden state h 'output by decoder at t' momentt`And the hidden state h output by the encoder at each momentiCalculating a score
Figure 622586DEST_PATH_IMAGE003
(ii) a After the score is processed by softmax, the score is processed by softmax
Figure 420646DEST_PATH_IMAGE004
Is converted into
Figure 711950DEST_PATH_IMAGE002
；
Obtaining a full-text thought vector c = { c = }t`}，t`∈T`。
6. The intelligence system modeling analysis method based on artificial intelligence algorithm of claim 5, characterized in that: the hidden state h 'output by the decoder at the t' momentt`And the hidden state h output by the encoder at each momentiCalculating a score
Figure 447825DEST_PATH_IMAGE003
Comprises the following steps:
Figure 115567DEST_PATH_IMAGE005
wherein the content of the first and second substances,
Figure 4019DEST_PATH_IMAGE006
、
Figure 466225DEST_PATH_IMAGE007
、
Figure 892658DEST_PATH_IMAGE008
representing learnable parameters, h ″t`And hiThe combined signals are input into a multi-layer perceptron to obtain a score, and tanh is an activation function.
7. The intelligence system modeling analysis method based on artificial intelligence algorithm of claim 1, characterized in that: the step of processing the n chapters based on the full-text thought vector c through the LSTM recurrent neural network to generate the chapter thought vector cp comprises the following steps:
the LSTM recurrent neural network comprises a trained encoder and a trained decoder;
analyzing n chapters based on the full-text thought vector c, wherein the n chapters contain K word vectors X, and the word vectors X = { X =kK belongs to K, K is an integer greater than or equal to 1, and xkIndicating input encoding at time k A word vector of the machine; when k =1, inputting a third initial hidden state g to the trained encoder0Sum word vector x1The encoder outputs the hidden shape g at this moment1(ii) a When k is>When 1, the hidden state g of the last time is input to the encoderk-1And the word vector x at this momentkThe encoder outputs the hidden state g at this momentk(ii) a Up to all word vectors xkAll input into the encoder, the hidden state g at the Kth moment is obtainedK；
Retiming, predicting word Y = { Y = }k`Inputting the data into a decoder in sequence, K' is an integer greater than or equal to 1, and yk`Represents the prediction value input to the decoder at time k'; when k =1, inputting a fourth initial hidden state g' into the trained decoder0And the initial predicted value y1The fourth initial latent state g ″0Hidden state g for encoder outputKInitial predicted value y1Is a custom value<bos>The decoder outputs the hidden state g' at this moment1And the predicted value y at the next time2(ii) a At k>At 1, the hidden state g' at the last moment is input to the decoderk`-1And the predicted value y at that timek`The decoder outputs the hidden state g' at this momentk`And the predicted value y at the next timek`-1(ii) a Until K' is reached, or until the decoder outputs the custom prediction value yk`+1=<eos>；
The chapter idea vector cp is generated from the hidden states output by the encoder and decoder.
8. The intelligence system modeling analysis method based on artificial intelligence algorithm of claim 7, wherein: the step of generating the chapter idea vector cp according to the hidden states output by the encoder and the decoder comprises the following steps: at time k' of the decoder, chapter idea vector cp may be generatedk`：
Figure 347779DEST_PATH_IMAGE009
Wherein, the first and the second end of the pipe are connected with each other,
Figure 340006DEST_PATH_IMAGE010
shows the chapter idea vector cp at time k `k`J is the jth moment of the encoder, gjA hidden state output for the j moment of the encoder;
weight of
Figure 238692DEST_PATH_IMAGE010
The calculation of (2):
hidden state g 'output by decoder at k' momentk`And the hidden state g output by the encoder at each momentjCalculating a score
Figure 949159DEST_PATH_IMAGE011
(ii) a After the score is processed by softmax, the score is processed by softmax
Figure 464364DEST_PATH_IMAGE012
Is converted into
Figure 576677DEST_PATH_IMAGE010
；
Get chapter idea vector cp = { cp = { (cp)k`}，k`∈K`。
9. The intelligence system modeling analysis method based on artificial intelligence algorithm of claim 8, characterized by: the hidden state g 'output at the k' th moment of the decoderk`And the hidden state g output by the encoder at each momentjCalculating a score
Figure 380684DEST_PATH_IMAGE011
The method comprises the following steps:
Figure 578448DEST_PATH_IMAGE013
wherein the content of the first and second substances,
Figure 375371DEST_PATH_IMAGE006
、
Figure 342190DEST_PATH_IMAGE007
、
Figure 582679DEST_PATH_IMAGE008
representing learnable parameters, g ″k`And gjThe combined signals are input into a multi-layer perceptron to obtain a score, and tanh is an activation function.
CN202210511565.0A 2022-05-12 2022-05-12 Information system modeling analysis method based on artificial intelligence algorithm Active CN114610871B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210511565.0A CN114610871B (en) 2022-05-12 2022-05-12 Information system modeling analysis method based on artificial intelligence algorithm

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210511565.0A CN114610871B (en) 2022-05-12 2022-05-12 Information system modeling analysis method based on artificial intelligence algorithm

Publications (2)

Publication Number Publication Date
CN114610871A true CN114610871A (en) 2022-06-10
CN114610871B CN114610871B (en) 2022-07-08

Family

ID=81870419

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210511565.0A Active CN114610871B (en) 2022-05-12 2022-05-12 Information system modeling analysis method based on artificial intelligence algorithm

Country Status (1)

Country Link
CN (1) CN114610871B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114970553A (en) * 2022-07-29 2022-08-30 北京道达天际科技股份有限公司 Information analysis method and device based on large-scale unmarked corpus and electronic equipment

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109726281A (en) * 2018-12-12 2019-05-07 Tcl集团股份有限公司 A kind of text snippet generation method, intelligent terminal and storage medium
CN110134782A (en) * 2019-05-14 2019-08-16 南京大学 A kind of text snippet model and Method for Automatic Text Summarization based on improved selection mechanism and LSTM variant
CN111651589A (en) * 2020-08-10 2020-09-11 中南民族大学 Two-stage text abstract generation method for long document
WO2020261234A1 (en) * 2019-06-28 2020-12-30 Tata Consultancy Services Limited System and method for sequence labeling using hierarchical capsule based neural network
CN113609284A (en) * 2021-08-02 2021-11-05 河南大学 Method and device for automatically generating text abstract fused with multivariate semantics

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109726281A (en) * 2018-12-12 2019-05-07 Tcl集团股份有限公司 A kind of text snippet generation method, intelligent terminal and storage medium
CN110134782A (en) * 2019-05-14 2019-08-16 南京大学 A kind of text snippet model and Method for Automatic Text Summarization based on improved selection mechanism and LSTM variant
WO2020261234A1 (en) * 2019-06-28 2020-12-30 Tata Consultancy Services Limited System and method for sequence labeling using hierarchical capsule based neural network
CN111651589A (en) * 2020-08-10 2020-09-11 中南民族大学 Two-stage text abstract generation method for long document
CN113609284A (en) * 2021-08-02 2021-11-05 河南大学 Method and device for automatically generating text abstract fused with multivariate semantics

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114970553A (en) * 2022-07-29 2022-08-30 北京道达天际科技股份有限公司 Information analysis method and device based on large-scale unmarked corpus and electronic equipment
CN114970553B (en) * 2022-07-29 2022-11-08 北京道达天际科技股份有限公司 Information analysis method and device based on large-scale unmarked corpus and electronic equipment

Also Published As

Publication number Publication date
CN114610871B (en) 2022-07-08

Similar Documents

Publication Publication Date Title
CN108984724B (en) Method for improving emotion classification accuracy of specific attributes by using high-dimensional representation
CN111897908B (en) Event extraction method and system integrating dependency information and pre-training language model
CN110765260A (en) Information recommendation method based on convolutional neural network and joint attention mechanism
CN112270379A (en) Training method of classification model, sample classification method, device and equipment
CN111159407A (en) Method, apparatus, device and medium for training entity recognition and relation classification model
CN111127146A (en) Information recommendation method and system based on convolutional neural network and noise reduction self-encoder
Kaur Incorporating sentimental analysis into development of a hybrid classification model: A comprehensive study
CN110851176B (en) Clone code detection method capable of automatically constructing and utilizing pseudo-clone corpus
CN113742733B (en) Method and device for extracting trigger words of reading and understanding vulnerability event and identifying vulnerability type
CN111831783B (en) Method for extracting chapter-level relation
CN112256866A (en) Text fine-grained emotion analysis method based on deep learning
CN113127716B (en) Emotion time sequence anomaly detection method based on saliency map
CN116484024A (en) Multi-level knowledge base construction method based on knowledge graph
CN113836896A (en) Patent text abstract generation method and device based on deep learning
CN114896386A (en) Film comment semantic emotion analysis method and system based on BilSTM
CN112818698A (en) Fine-grained user comment sentiment analysis method based on dual-channel model
CN110569355B (en) Viewpoint target extraction and target emotion classification combined method and system based on word blocks
CN114610871B (en) Information system modeling analysis method based on artificial intelligence algorithm
CN109979461A (en) A kind of voice translation method and device
CN117350898A (en) Intelligent early warning system and method for annual patent fee
CN112434512A (en) New word determining method and device in combination with context
CN111309849A (en) Fine-grained numerical information extraction method based on joint learning model
CN115630223A (en) Service recommendation method and system based on multi-model fusion
CN113297385A (en) Multi-label text classification model and classification method based on improved GraphRNN
CN113255360A (en) Document rating method and device based on hierarchical self-attention network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CP01 Change in the name or title of a patent holder
CP01 Change in the name or title of a patent holder

Address after: 100085 room 703, 7 / F, block C, 8 malianwa North Road, Haidian District, Beijing

Patentee after: Beijing daoda Tianji Technology Co.,Ltd.

Address before: 100085 room 703, 7 / F, block C, 8 malianwa North Road, Haidian District, Beijing

Patentee before: Beijing daoda Tianji Technology Co.,Ltd.